On March 24, roughly 1,500 investors will file into Y Combinator's Demo Day to watch what may be the accelerator's largest class ever pitch their startups. The exact headcount remains something of a moving target—third-party trackers place it somewhere between 161 and 196 companies, depending on who's counting and when—but the composition is unmistakable. This batch is drowning in AI agents. Not the helpful-assistant kind. The replace-entire-departments kind.
Scroll through the public company directory or the steady drip of Launch HN posts that dominated March, and a pattern emerges: startup after startup promising to automate work in industries that have spent decades successfully ignoring Silicon Valley. Healthcare operations. Legal compliance. Satellite mission assurance. Commercial insurance brokerage. Even hotel management. It's as if every founder looked at the same thesis—"LLMs can finally crack legacy workflows"—and decided to test it in a different vertical.
Among them: François Chollet, the engineer behind Keras, who's launching a new venture called Ndea with co-founder Mike Knoop. Or the team at Constellation Space, a cluster of former SpaceX, NASA, and Blue Origin engineers now building AI tools for satellite operators and already running design-partner tests. The pedigree alone suggests this isn't amateur hour. What's less clear is whether the technology has caught up to the ambition.
The Numbers Game
YC itself hasn't published an official roster, pointing investors instead to its Startup Directory and Launch YC platform rather than a tidy PDF. That's left the counting to outsiders. VC Corner, which cross-references the directory with LinkedIn and Crunchbase, tallied 196 startups in a database published March 9. An independent dashboard parsing YC's public Algolia API offers sector breakdowns but dodges a final number. The ambiguity may be deliberate—YC has always preferred flexible storytelling over static lists—but it does make comparative analysis harder than it should be.
What isn't up for debate is the deal structure. Since expanding to four cohorts annually in 2024, YC invests $500,000 per company: $125,000 via a post-money SAFE for 7% equity, plus $375,000 through an uncapped MFN SAFE, alongside pro-rata rights. Demo Day remains invite-only, a tradition that's held even as batch sizes have crept upward. Roughly 1,500 people will be in the room. Perhaps more than some founders expected when they applied.
Agents Everywhere
The most arresting thing about W26 isn't the size. It's the sheer concentration of companies selling AI not as a tool but as a replacement for human jobs. Eos AI bills itself as an "autonomous operating system for healthcare," indexing datasets measured in petabytes. Beacon Health is deploying agents for primary care. CellType? It calls itself an "agentic drug company," a phrase that would've sounded like science fiction five years ago.
Legal and compliance teams are getting similar treatment. Veriad offers what it describes as AI compliance officers for marketing departments. Oxus automates internal audits for companies eyeing IPOs or already public. Maywood handles deal execution for investment banks, the kind of grunt work that's kept junior analysts employed for generations.
Finance, unsurprisingly, isn't immune. Panta is tackling commercial insurance brokerage with AI agents. FullSeam pitches itself as an "AI employee" for accounting, which raises the uncomfortable question of what happens to the human employees currently doing that work.
Even niche sectors are getting swept up. Lance claims to support more than 50 hotels with autonomous agents managing operations. Corvera, which reportedly raised $2 million within its first month according to its YC profile, is targeting retail operations. BeeSafe AI—founded by PhD researchers from Google and Microsoft—takes a stranger approach: it engages scammers directly, mining verified intelligence on fraud networks. One imagines the pitch meetings were lively.
Hard Tech Makes a Comeback
Buried in the agent frenzy is a quieter subplot: YC is backing hard-tech and defense startups that would've seemed wildly out of place a decade ago, back when the accelerator was best known for consumer apps and SaaS tools.
Voltair is deploying a globally distributed autonomous drone network for Earth observation, with power utilities as its initial customer base. Terranox AI is using machine learning for uranium exploration—yes, uranium. HLabs is building a domestic supply chain for robotics manufacturing, a bet that geopolitics matters as much as technology.
Then there's Constellation Space. The company stands out not just for its founder pedigree but for the problem it's solving: mission assurance for satellite operators. As commercial space scales—and it is scaling fast—the bottleneck isn't launch capacity anymore. It's making sure satellites work once they're up there. The fact that Constellation is already testing with design partners suggests the pain point is real, not speculative.
The Picks and Shovels
Beneath all those vertical applications, a handful of W26 companies are building infrastructure specifically for the AI agent boom. Sentrial offers production monitoring for agents, addressing the observability gap that emerges when you let software make decisions autonomously. RunAnywhere is focused on on-device inference, targeting Apple Silicon and mobile SDKs. Zatanna is doing something more conceptually interesting: turning legacy software into agent-friendly APIs, essentially teaching decades-old systems to speak the language of LLMs.
Even adjacent categories are pivoting toward AI. Booko applies dynamic pricing to time-slot businesses—escape rooms, spa appointments, that sort of thing—using algorithms that weren't economical at small scale before modern machine learning made them cheap. Byteport is tackling global upload acceleration for files ranging from 1GB to 100TB, infrastructure that matters more as training datasets grow unwieldy.
Early Money
Fundraising announcements have been trickling out throughout the batch. Patientdesk AI, which automates dental front desks, announced a $1 million pre-seed on February 19. Corvera's self-reported $2 million raise came fast, though the company hasn't disclosed timing or lead investors publicly. Eight Capital's weekly "YC W26 Seed Spotlight" newsletters have tracked fundraising progress through March, but no comprehensive total has emerged ahead of Demo Day.
Gauging how much capital the cohort will ultimately attract is trickier. YC's 2024 Top Companies list, published that April, showed alumni generating $57.2 billion in revenue across featured companies in 2023, with a combined valuation of $458 billion and more than 80,000 employees. Of course, those figures reflect companies many years removed from their batch, not fresh graduates. But they do set a benchmark—or at least an aspiration—for what a tiny fraction of any cohort might eventually deliver.
What It All Means
TechCrunch has noted strong AI representation in YC's recent cohorts, including Winter 2024 and Summer 2024. But W26 feels different. This isn't about building better chatbots or productivity tools. It's about rebuilding entire workflows in law, healthcare, finance, and logistics around the assumption that software can now handle tasks requiring human judgment.
That's a big assumption. Whether it holds at scale is the question investors will be wrestling with on March 24.
The batch carries enough founder credibility—Chollet's track record with Keras, the SpaceX alumni, the PhD researchers—to suggest the technical challenges aren't trivial. And the relentless focus on legacy industries suggests founders believe the biggest opportunities lie not in inventing new categories but in finally automating the old, boring ones.
Demo Day will reveal which bets resonate with checkwriters. But the composition of the batch already reveals what YC believes about where AI is headed: straight into the regulated, high-margin, decidedly unsexy industries that have spent decades successfully ignoring Silicon Valley's disruption playbook. Whether those industries are finally ready to be disrupted—or whether they'll resist one more time—is another matter entirely.
